A method for monitoring plateau meadow baldness and degradation degree based on improved YOLOv8

Abstract

The present application relates to a kind of plateau meadow alopecia and degradation degree monitoring method based on improved YOLOv8, comprising: constructing plateau meadow alopecia data set, and pretreatment;Wherein, the plateau meadow alopecia data set includes: healthy meadow, mild alopecia, three kinds of ground data sets of severe bareness class;The backbone network and neck network of YOLOv8 model are improved, and PMMINC-YOLOv8 model is constructed;The PMMINC-YOLOv8 model is trained using the data set after pretreatment, and monitoring model is acquired;Plateau meadow alopecia and degradation degree monitoring are carried out using the monitoring model.The present application can realize the high-precision, automatic monitoring of plateau meadow degradation condition.

Description

Technical Field

[0001] This invention relates to the field of grassland ecological remote sensing monitoring and computer vision image processing technology, and in particular to a method for monitoring patchy baldness and degradation of alpine meadows based on an improved YOLOv8. Background Technology

[0002] Alpine meadows are important ecological barriers and livestock bases in my country. However, affected by climate change, rodent infestations, overgrazing, and human factors, these meadows have experienced varying degrees of degradation, forming typical "bald patches" of bare land (such as the early signs of black soil degradation). If not addressed promptly, these bare patches will rapidly expand, leading to severe soil erosion. Therefore, timely and accurate monitoring of the location and area of ​​meadow bare patches is crucial for grassland ecological restoration.

[0003] Current monitoring methods mainly have the following problems: Traditional manual ground surveys: The harsh environment of the plateau makes manual patrols extremely inefficient and difficult to accurately count the area of ​​bald patches, which cannot meet the needs of large-scale monitoring.

[0004] Limitations of existing deep learning models (such as YOLOv5, YOLOv8n, etc.): Significant background interference: The background of high-altitude meadows is complex, containing numerous gravel, rodent burrows, and cloud shadows, making them prone to misidentification as bald patches. Difficulty in feature differentiation: For areas of "mild baldness" (sparse vegetation but not completely bare), the texture features fall between those of healthy grassland and bare soil. Existing models lack sufficient feature extraction capabilities, easily leading to missed detections or misclassifications. Significant scale differences: Bald patches vary in size, ranging from rodent-infested patches tens of centimeters in diameter to continuous bare land hundreds of meters in length. Existing algorithms lack effective cross-scale feature fusion mechanisms. Summary of the Invention

[0005] The purpose of this invention is to propose a monitoring method for bald patches and degradation of alpine meadows based on an improved YOLOv8. This method aims to solve the problems of low accuracy in identifying bald patches (especially mild bald patches with blurred edges) in complex backgrounds and susceptibility to interference from gravel shadows in existing technologies, and to achieve high-precision and automated monitoring of the degradation status of alpine meadows.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for monitoring patchy alopecia and degradation in high-altitude meadows based on an improved YOLOv8 system includes: A high-altitude meadow alopecia areata dataset was constructed and preprocessed; the high-altitude meadow alopecia areata dataset includes three categories of surface datasets: healthy meadow, mild alopecia areata, and severe bare ground. The backbone and neck network of the YOLOv8 model were improved to construct the PMMINC-YOLOv8 model; The PMMINC-YOLOv8 model was trained using the preprocessed dataset to obtain the monitoring model; The monitoring model was used to monitor patchy alopecia and the degree of degradation in alpine meadows.

[0007] Optionally, a dataset of patchy alopecia in the alpine meadows is constructed and preprocessed, including: Orthophoto data of alpine meadows were collected, and the image data were cropped and enhanced, with alopecia areata areas marked.

[0008] Optionally, improvements to the backbone network of the YOLOv8 model include: In the backbone network, an ADown downsampling module is introduced to replace the stride convolution of the YOLOv8 model in order to distinguish between soil and vegetation textures. The ADown downsampling module uses parallel max pooling branches and convolution branches to extract the texture difference features between soil and vegetation while reducing the resolution.

[0009] Optionally, improvements to the neck network of the YOLOv8 model include: A REPC3 feature fusion module is introduced into the neck network to adapt to multi-scale alopecia areata; the REPC3 feature fusion module captures the contour semantic information of large-scale alopecia areata and the detailed texture information of small-scale alopecia areata simultaneously during the feature fusion stage.

[0010] Optionally, training the PMMINC-YOLOv8 model using the preprocessed dataset includes: The preprocessed dataset is input into the PMMINC-YOLOv8 model to obtain the prediction results; The prediction results are evaluated and the model is optimized using the Slide Loss function.

[0011] Optionally, inputting the preprocessed dataset into the PMMINC-YOLOv8 model includes: The processed data is represented digitally as a three-dimensional tensor; wherein the dimensions of the three-dimensional tensor are H×W×3, corresponding to the height, width and RGB channels of the image, respectively. After the three-dimensional tensor is input into the backbone network, it is processed by the parallel max pooling branch and convolution branch in the ADown downsampling module. The feature maps obtained after processing the two branches are then concatenated in the channel dimension to output feature maps of different scales. After feature maps of different scales are input into the neck network, feature fusion is performed through the REPC3 feature fusion module. The data enters the REPC3 feature fusion module and is processed in parallel through three paths: 3x3 convolution, 1x1 convolution and direct identity mapping. The results of these three branches are then concatenated together along the channel dimension to complete feature fusion.

[0012] Optionally, the max-pooling branch is designed to extract high-response feature points in the bare soil region and filter out the shadow background; The convolution branch extracts the texture difference features between soil and vegetation through a convolution operation with a preset stride, and then splices them to generate a downsampled feature map.

[0013] Optionally, the Slide Loss function increases the loss weight for predicted bounding boxes with an IoU value lower than a preset value.

[0014] Optionally, using the monitoring model to monitor patchy alopecia and the degree of degradation in alpine meadows includes: Monitor the center coordinates, length and width dimensions, and degradation level labels of bald patches.

[0015] The beneficial effects of this invention are as follows: High recognition accuracy: Effectively solves the problems of blurred edges and background interference in meadow bald patches. The recognition accuracy of "bald patches" is significantly improved, and it can accurately distinguish between sparse vegetation and bare soil.

[0016] Strong anti-interference capability: Through feature extraction by the ADown module, it can effectively eliminate false detections caused by common high-altitude phenomena such as gravel, yak dung, and cloud shadows.

[0017] Multi-scale monitoring: The REPC3 module can simultaneously and accurately detect both small, fragmented rodent-infested bald patches and large areas of degraded bare land.

[0018] Lightweight and efficient: The improved model has fewer parameters and faster inference speed, making it suitable for embedded devices in high-altitude inspection drones. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of a method for monitoring patchy alopecia and degradation in alpine meadows based on an improved YOLOv8 according to an embodiment of the present invention. Figure 2This is a schematic diagram illustrating the process of processing image data using the PMMINC-YOLOv8 model according to an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 and Figure 2 As shown, this embodiment proposes a method for monitoring patchy alopecia and degradation in alpine meadows based on an improved YOLOv8, including: A high-altitude meadow alopecia areata dataset was constructed and preprocessed; the high-altitude meadow alopecia areata dataset includes three categories of surface datasets: healthy meadow, mild alopecia areata, and severe bare ground. The backbone and neck network of the YOLOv8 model were improved to construct the PMMINC-YOLOv8 model; The PMMINC-YOLOv8 model was trained using the preprocessed dataset to obtain the monitoring model; The monitoring model was used to monitor patchy alopecia and the degree of degradation in alpine meadows.

[0024] Furthermore, a dataset of patchy alopecia areata in the high-altitude meadows was constructed and preprocessed, including: Orthophoto data of alpine meadows were collected, and the image data were cropped and enhanced, with alopecia areata areas marked.

[0025] Specifically, in this embodiment, an orthophotos of alpine meadows are collected using a drone, and the images are preprocessed and labeled.

[0026] Furthermore, improvements to the backbone network of the YOLOv8 model include: In the backbone network, an ADown downsampling module is introduced to replace the stride convolution of the YOLOv8 model in order to distinguish between soil and vegetation textures. The ADown downsampling module uses parallel max pooling branches and convolution branches to extract the texture difference features between soil and vegetation while reducing the resolution.

[0027] Improvements to the neck network of the YOLOv8 model include: A REPC3 feature fusion module is introduced into the neck network to adapt to multi-scale alopecia areata; the REPC3 feature fusion module captures the contour semantic information of large-scale alopecia areata and the detailed texture information of small-scale alopecia areata simultaneously during the feature fusion stage.

[0028] Specifically, in this embodiment, the improved monitoring model P-YOLOv8 is constructed based on the YOLOv8n network, and the following improvements are made: (1) Backbone improvement: The ADown downsampling module is introduced. The ADown module uses parallel max pooling and convolution branches to extract the texture difference features of soil and vegetation (such as the contrast between yellowish-brown soil texture and green grass texture) while reducing the resolution, effectively filtering high-frequency background noise such as gravel and shadows.

[0029] (2) Neck network improvement: The REPC3 (Replicated Convolutional Cross-Scale Feature Fusion) feature fusion module is introduced. Considering the coexistence of bald patches ranging from tiny "rat hole-level" patches to large areas of bare land, this module captures both the contour semantic information of large-scale bald patches and the detailed texture information of small-scale bald patches during the feature fusion stage.

[0030] (3) Improved loss function: The Slide Loss function is adopted. For the edge of meadow bald patches, which often presents a blurred transition state of "mixed grass and soil" (difficult sample), the penalty weight for blurred boundary samples is increased by using the IoU weighting strategy, thereby improving the model's segmentation and localization accuracy of bald patch edges.

[0031] Furthermore, training the PMMINC-YOLOv8 model using the preprocessed dataset includes: The preprocessed dataset is input into the PMMINC-YOLOv8 model to obtain the prediction results; The prediction results are evaluated and the model is optimized using the Slide Loss function.

[0032] Specifically, in this embodiment, the pre-divided training set is input into the constructed PMMINC-YOLOv8 model for iterative training. The Slide Loss is used to optimize the network weights, enabling the model to learn to distinguish between "real alopecia areata" and "false background".

[0033] The model inputs the meadow orthophoto of the area to be monitored into the trained model and outputs monitoring results including the center coordinates of the bald patch, the coverage area bounding box, and the degradation level category.

[0034] The following example uses a typical degraded meadow area on the Qinghai-Tibet Plateau to illustrate the method of this embodiment: Dataset Construction: A typical degraded meadow area on the Qinghai-Tibet Plateau was selected as the study area. Orthophotos were collected using drones. "Alopecia Areas" were labeled.

[0035] The specific processing flow of the P-YOLOv8 model for a single meadow image: This embodiment will take a meadow orthophoto extracted from the dataset of Embodiment 1 as an example to illustrate in detail how the P-YOLOv8 model proposed in this invention works in concert through its three core improvement modules (ADown, REPC3, and Slide Loss) to ultimately achieve high-precision detection of "bald patches".

[0036] Input and feature digitization definition: First, the meadow orthophoto image is fed into the P-YOLOv8 model. Digitally, the image is represented as a three-dimensional tensor I with dimensions H × W × 3, corresponding to the image's height, width, and the three RGB (red, green, and blue) color channels, respectively. These RGB values ​​directly reflect the physical characteristics of the land surface and form the basis for the model's judgments. High G channel value regions: appear green in the image, corresponding to healthy meadow coverage areas, and are regarded as "background" that needs to be suppressed in the algorithm.

[0037] High R and B mixed value areas appear as yellowish-brown in the image, corresponding to the "bald patches" that need to be detected, i.e., exposed soil or sandy surfaces.

[0038] Feature extraction and denoising of the backbone network (core improvement: ADown module): After the image is input into the backbone network, at key nodes where feature map resolution needs to be reduced (downsampling), this invention replaces the strided convolution of traditional YOLOv8 with the ADown module. Its specific workflow is as follows: Parallel dual-branch processing: When a feature map F_in representing local surface texture arrives at the ADown module, it is simultaneously fed into two parallel processing branches: Branch A (MaxPool): Through max pooling operation, it aims to extract and retain the bright pixels in the center of the "bald patch" area caused by the high reflectivity of the soil, while effectively filtering out low-response background noise such as shadows in the grass and small stones.

[0039] Branch B (Conv - Convolutional Branch): Through a convolution operation with a stride of 2, the weights of its convolution kernel are optimized after training to specifically recognize specific textures, focusing on learning and extracting the significant differences between the high-frequency textures of "gravelly and sandy textures of yellowish-brown bare soil" and "linear and fibrous textures of green grass".

[0040] Feature concatenation and output: The feature maps obtained after processing the two branches are concatenated along the channel dimension. The concatenated output feature map F_out retains the salience information of the bald patches and enhances the edge texture contrast with the vegetation background.

[0041] Cross-scale feature fusion in the neck network (core improvement: REPC3 module): The neck network receives feature maps X at different depths (i.e., different scales) output by the backbone network and performs efficient fusion using the REPC3 module: Multi-scale information fusion: The REPC3 module, through its internal multi-branch structure of 3x3 convolutions, 1x1 convolutions, and identity mappings, can simultaneously process features from different levels. Deep feature maps provide global semantic information such as the contours and shapes of large, contiguous bare areas, while shallow feature maps provide local details such as the color and texture of small, scattered bald patches.

[0042] Unified Feature Representation: By aggregating this multi-scale and multi-dimensional information, the REPC3 module can construct a unified feature representation with high responsiveness to a range of bald patches ranging from tens of centimeters to hundreds of meters of bare land, thus effectively solving the problem of huge differences in target scale.

[0043] Loss calculation and boundary optimization during model training (core improvement: Slide Loss): During the model training phase, Slide Loss is used to evaluate and optimize the prediction results, specifically addressing the problem of locating blurred edges in alopecia areata: Hard sample identification: The model predicts bounding boxes based on feature maps and calculates the intersection-over-union (IoU) with the ground truth labeled boxes. For patchy bare land with blurred boundaries due to sparse vegetation and mixed grass and soil, the IoU value of the predicted boxes is usually low, and they are identified as "hard samples".

[0044] Dynamic loss weighting: The core mechanism of Slide Loss is that it dynamically weights the data based on the IoU value, automatically increasing the proportion of these "hard samples" in the total loss.

[0045] Focused optimization: This weighted mechanism forces the model to invest more learning resources in studying how to accurately define those fuzzy and difficult-to-locate grass-soil mixture boundaries when updating parameters through backpropagation, thereby significantly improving the overall localization accuracy of the model.

[0046] Head calculation and final result output: After sufficient training, the model can be used for actual monitoring. When a new meadow image is input, after the above processing steps, the detection head finally completes the calculation and outputs a monitoring result set O. Each element in set O is a tuple of the form (xc, yc, w, h, conf), where: (xc, yc): the center pixel coordinates of a "bald patch" detected in the original image; (w, h): the coverage area of ​​the bald patch (width and height of the bounding box); and conf (Confidence): a confidence score representing the model's degree of confidence in the judgment that "the area within the bounding box is identified as bald patches."

[0047] Through this detailed implementation process, the P-YOLOv8 model ultimately achieved high-precision, automated localization and bounding of all "bald patches" in the alpine meadow imagery. Example 3: Effect Verification Experiments show that the method achieves an average precision (mAP) of 84.1% on the alpine meadow test set, accurately identifying tiny bald patches with a diameter of only 20 cm, and does not misclassify grayish-white rocks as bare soil, thus verifying its robustness in complex alpine environments.

[0048] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for monitoring patchy alopecia and degradation in high-altitude meadows based on an improved YOLOv8 system, characterized in that, include: A high-altitude meadow alopecia areata dataset was constructed and preprocessed; the high-altitude meadow alopecia areata dataset includes three categories of surface datasets: healthy meadow, mild alopecia areata, and severe bare ground. The backbone and neck network of the YOLOv8 model were improved to construct the PMMINC-YOLOv8 model; The PMMINC-YOLOv8 model was trained using the preprocessed dataset to obtain the monitoring model; The monitoring model was used to monitor patchy alopecia and the degree of degradation in alpine meadows.

2. The method for monitoring patchy alopecia and degradation in alpine meadows based on improved YOLOv8 according to claim 1, characterized in that, A dataset of patchy alopecia in the high-altitude meadows was constructed and preprocessed, including: Orthophoto data of alpine meadows were collected, and the image data were cropped and enhanced, with alopecia areata areas marked.

3. The method for monitoring patchy alopecia and degradation in alpine meadows based on improved YOLOv8 according to claim 1, characterized in that, The improvements to the backbone network of the YOLOv8 model include: In the backbone network, an ADown downsampling module is introduced to replace the stride convolution of the YOLOv8 model in order to distinguish between soil and vegetation textures. The ADown downsampling module uses parallel max pooling branches and convolution branches to extract the texture difference features between soil and vegetation while reducing the resolution.

4. The method for monitoring patchy alopecia and degradation in alpine meadows based on improved YOLOv8 according to claim 1, characterized in that, Improvements to the neck network of the YOLOv8 model include: A REPC3 feature fusion module is introduced into the neck network to adapt to multi-scale alopecia areata; the REPC3 feature fusion module captures the contour semantic information of large-scale alopecia areata and the detailed texture information of small-scale alopecia areata simultaneously during the feature fusion stage.

5. The method for monitoring patchy alopecia and degradation in alpine meadows based on improved YOLOv8 according to claim 1, characterized in that, Training the PMMINC-YOLOv8 model using the preprocessed dataset includes: The preprocessed dataset is input into the PMMINC-YOLOv8 model to obtain the prediction results; The prediction results are evaluated and the model is optimized using the Slide Loss function.

6. The method for monitoring patchy alopecia and degradation in high-altitude meadows based on improved YOLOv8 according to claim 5, characterized in that, Inputting the preprocessed dataset into the PMMINC-YOLOv8 model includes: The processed data is represented digitally as a three-dimensional tensor; wherein the dimensions of the three-dimensional tensor are H × W × 3, corresponding to the height, width and RGB channels of the image, respectively. After the three-dimensional tensor is input into the backbone network, it is processed by the parallel max pooling branch and convolution branch in the ADown downsampling module. The feature maps obtained after processing the two branches are then concatenated in the channel dimension to output feature maps of different scales. After feature maps of different scales are input into the neck network, feature fusion is performed through the REPC3 feature fusion module. The data enters the REPC3 feature fusion module and is processed in parallel through three paths: 3x3 convolution, 1x1 convolution and direct identity mapping. The results of these three branches are then concatenated together along the channel dimension to complete feature fusion.

7. The method for monitoring patchy alopecia and degradation in alpine meadows based on improved YOLOv8 according to claim 6, characterized in that, The max-pooling branch is designed to extract high-response feature points in the bare soil region and filter out the shadow background. The convolution branch extracts the texture difference features between soil and vegetation through a convolution operation with a preset stride, and then splices them to generate a downsampled feature map.

8. The method for monitoring patchy alopecia and degradation in alpine meadows based on improved YOLOv8 according to claim 5, characterized in that, The Slide Loss function increases the loss weight for predicted bounding boxes whose IoU value is lower than a preset value.

9. The method for monitoring patchy alopecia and degradation in alpine meadows based on improved YOLOv8 according to claim 1, characterized in that, Monitoring of patchy alopecia and degradation in alpine meadows using the aforementioned monitoring model includes: Monitor the center coordinates, length and width dimensions, and degradation level labels of bald patches.

Images